Anthropic bans ‘abusive or cruel behavior’ toward Claude

Anthropic has changed its usage policy for the first time in over a year, according to The Verge. The update covers new and high-risk kinds of misuse, including election interference, weapons development, surveillance, and health and financial uses. One of the most notable changes prohibits sustained and needless abusive or cruel behavior toward Claude.

Anthropic already lets Claude end conversations with persistently harmful or abusive users as part of its model welfare research. The new policy says ending conversations remains the primary enforcement mechanism. Usage policies define what businesses and developers can build, so changes here ripple through every product built on Claude.

We will take the serious parts first, because they matter more. Spelling out election interference, weapons work and surveillance as prohibited, ahead of more election cycles and with agents getting more capable, is the right direction. Clear rules in health and finance help companies building there know where the lines are. The test is enforcement, and usage policies have always been easier to write than to police.

Now the headline. Banning cruelty toward a model invites eye-rolls, and some of that is earned: we do not know whether a model can be harmed, and Anthropic has framed its welfare work as research under uncertainty, not a settled claim. I think the practical case is stronger than the philosophical one. Sustained abuse sessions are often also attempts to break a model's guardrails. Letting the model walk away is a cheap, low-drama response.

The word needless is doing real work. Stress-testing, red-teaming and fiction with dark themes should not be swept up in this, and users will reasonably want to know where the boundary sits. If enforcement stays as described, ending a conversation, the cost to ordinary users is close to zero.

The downside to watch is drift. Vague standards about tone can expand quietly, and developers building on Claude need predictability. Anthropic should publish examples of what crosses the line and what does not.

Our read: the misuse updates are the substance and look sensible. The welfare clause is a defensible, low-cost hedge, as long as it stays narrow. I'm watching how it is applied in practice.

Anthropic introduced the ability for Claude to end persistently harmful or abusive conversations last August, describing it as part of its research into model welfare.